IMASS: An Intelligent Microblog Analysis and Summarization System Jui-Yu Weng Cheng-Lun Yang Bo-Nian Chen Yen-Kai Wang Shou-De Lin Department of Computer Science and Information Engineering National Taiwan University {r98922060,r99944042,f92025,b97081,sdlin}@csie.ntu.edu.tw methods can hardly be applied to Microblog texts Abstract because many posts/responses contain only one sentence. This paper presents a system to summarize Below we first describe some special characte- a Microblog post and its responses with the ristics that deviates the Microblog summarization goal to provide readers a more constructive task from general text summarization. and concise set of information for efficient a. The number of sentences is limited, and sen- digestion. We introduce a novel two-phase tences are usually too short and casual to con- summarization scheme. In the first phase, tain sufficient structural information or cue the post plus its responses are classified in- phrases. Unlike normal blogs, there is a strict to four categories based on the intention, limitation on the number of characters for each interrogation, sharing, discussion and chat. post (e.g. 140 characters for Twitter and Plurk For each type of post, in the second phase, maximum). Microblog messages cannot be we exploit different strategies, including treated as complete documents so that we can- opinion analysis, response pair identifica- not take advantage of the structural information. tion, and response relevancy detection, to Furthermore, users tend to regard Microblog as summarize and highlight critical informa- a chatting board. They write casually with tion to display. This system provides an al- slangs, jargons, and incorrect grammar. ternative thinking about machine- b. Microblog posts can serve several different summarization: by utilizing AI approaches, purposes. At least three different types of posts computers are capable of constructing dee- are observed in Microblogs, expressing feeling, per and more user-friendly abstraction. sharing information, and asking questions. Structured language is not the only means to 1 Introduction achieve those goals. For example, people sometimes use attachment, as links or files, for As Microblog services such as Twitter have be- sharing, and utilize emoticons and pre-defined come increasingly popular, it is critical to re- qualifiers to express their feelings. The diver- consider the applicability of the existing NLP sity of content differ Microblogs from general technologies on this new media sources. Take news articles. Consequently, using one mold to summarization for example, a Microblog user fit all types of Microblog posts is not sufficient. usually has to browse through tens or even hun- Different summarization schemes for posts dreds of posts together with their responses daily, with different purposes are preferred. therefore it can be beneficial if there is an intelli- c. Posts and responses in Microblogs are more gent tool assisting summarizing those information. similar to a multi-persons dialogue corpus. One Automatic text summarization (ATS) has been of the main purposes of a Microblog is to serve investigated for over fifty years, but the majority of as the fast but not instant communication the existing techniques might not be appropriate channel among multiple users. Due to the free- for Microblog write-ups. For instance, a popular chatting, multi-user characteristics, the topic of kind of approaches for summarization tries to iden- a post/response thread can drift quickly. Some- tify a subset of information, usually in sentence times, the topic of discussion at the end of the form, from longer pieces of writings as summary thread is totally unrelated to that of the post. (Das and Martins, 2007). Such extraction-based 133 Proceedings of the ACL-HLT 2011 System Demonstrations, pages 133–138, Portland, Oregon, USA, 21 June 2011. c 2011 Association for Computational Linguistics This paper introduces a framework that summariz- kind of posts is to find out which replies really re- es a post with its responses. Motivated by the ab- spond to the question. We created a response re- ovementioned characteristics of Microblogs, we levance detection component to serve as its plan to use a two-phase summarization scheme to summarization mechanism. develop different summarization strategies for dif- The Sharing posts are very frequently observed ferent type of posts (see Figure 1). In the first in Microblog as Microbloggers like to share inter- phase, a post will be automatically classified into esting websites, pictures, and videos with their several categories including interrogation, discus- friends. Other people usually write down their sion, sharing and chat based on the intention of the comments or feelings on the shared subjects in the users. In the second phase, the system chooses dif- responses. To summarize such posts, we obtain the ferent summarization components for different statistics on how many people have positive, neu- types of posts. tral, and negative attitude toward the shared sub- The novelties of this system are listed below. jects. We introduce the opinion analysis 1. Strategically, we propose an underlying 2-phase component that provides the analysis on whether framework for summarizing Microblog posts. the information shared is recommended by the res- The system can be accessed online at pondents. http://mslab.csie.ntu.edu.tw/~fishyz/plurk/. We also observe that some posts contain charac- 2. Tactically, we argue that it is possible to inte- teristics of both Interrogation and Sharing. The grate post-intention classification, opinion anal- users may share a hyperlink and ask for others’ ysis, response relevancy and response-pair opinions at the same time. We create a category mining to create an intelligent summarization named Discussion for these posts, and apply both framework for Microblog posts and responses. response ranking and opinion analysis engines on We also found that the content features are not this type of posts. as useful as the temporal or positional features Finally, there are posts which simply act as the for text mining in Microblog. solicitation for further chat. For example, one user 3. Our work provides an alternative thinking about writes “so sad…” and another replies “what hap- ATS. It is possible to go beyond the literal pened”. We name this type of posts/responses as meaning of summarization to exploit advanced Chat. This kind of posts can sometimes involve text mining methods to improve the quality and multiple persons and the topic may gradually drift usability of a summarization system. to a different one. We believe the plausible sum- marization strategy is to group different messages 2 Summarization Framework and Expe- based on their topics. Therefore for Chat posts, we riments designed a response pair identification system to accomplish such goal. We group the related res- Below we discuss our two-phase summarization ponses together for display, and the number of framework and the experiment results on each in- groups represents the number of different topics in dividual component. Note that our experiments this thread. were tested on the Plurk dataset, which is one of Figure 1 shows the flow of our summarization the most popular micro-blogging platforms in Asia. Our observation is that Microblog posts can have different purposes. We divide them into four categories, Interrogation, Sharing, Discussion, and Chat. The Interrogation posts are questions asked in public with the hope to obtain some useful answers from friends or other users. However, it is very common that some repliers do not provide mea- ningful answers. The responses might serve the purpose for clarification or, even worse, have noth- ing to do with the question. Hence we believe the most appropriate summarization process for this Figure 1. System architecture 134 framework. When an input post with responses Here we design a two-level classification comes in, the system first determines its intention, framework using Naïve-Bayes classifiers which based on which the system adopts proper strategies takes advantage of the learned 6-gram language for summarization. Below we discuss the technical model probabilities as features. First of all, we parts of each sub-system with experiment results. train a binary classifier to determine if a post or a reply is opinionative. This step is called the subjec- 2.1 Post Intention Classification tivity test. If the answer is yes, we then use another This stage aims to classify each post into four cat- binary classifier to decide if the opinion is positive egories, Interrogation, Sharing, Discussion, and or negative. The second step is called the polarity Chat. One tricky issue is that the Discussion label test. is essentially a combination of interrogation and For subjectivity test, we manually annotate 3244 sharing labels. Therefore, simply treating it as an posts, in which half is subjective and half is objec- independent label and use a typical multi-label tive. The 10-fold cross validation shows average learning method can hurt the performance. We ob- accuracy of 70.5%. tain 76.7% (10-fold cross validation) in accuracy For polarity test, we exploit the built-in emoti- by training a four-class classifier using the 6-gram cons in Plurk to automatically extract posts with character language model. To improve the perfor- positive and negative opinions. We collect 10,000 mance, we design a decision-tree based framework positive and 10,000 negative posts as training data that utilizes both manually-designed rules and dis- to train a language model of Naïve Bayes classifier, criminant classification engine (see Figure 2). The and evaluate on manually annotated data of 3121 system first checks whether the posts contains posts, with 1624 positive and 1497 negative to ob- URLs or pointers to files, then uses a binary clas- tain accuracy of 0.722. sifier to determine whether the post is interrogative. 2.3 Response Pair Identification For the experiment, we manually annotate 6000 posts consisting of 1840 interrogation, 2002 shar- Conversation in micro-blogs tends to diverge into ing, 1905 chat, and 254 discussion posts.